Number Class
← Back to Explore
Computer Science Artificial Intelligence & Data Natural Language Processing

Search Engines Never Actually Learned What Words Mean

Google's own 2019 update touched just 1 in 10 queries — and production systems still run 1988-era keyword matching alongside every neural embedding.

S.J. Nam 8 min read

A paper that would eventually reshape how billions of searches work every day was, at first, rejected. In 2013, Tomas Mikolov and his colleagues at Google submitted the work that introduced word2vec to the ICLR conference — and reviewers turned it down. It then took months before Google would even let the code be released to the public. That is not the origin story most explainers tell about semantic search. The version usually told is triumphant: keyword search was primitive, meaning-aware search is smarter, and the machines finally learned to understand us. The real story is messier, slower, and more interesting — and it points to a conclusion the industry doesn't advertise much: semantic search never actually replaced keyword matching. It exposed how much of "understanding" was geometry all along, and the field's own engineering practice has quietly walked back toward hybrids of old and new.

The Paper Nobody Wanted

Start with the rejection itself, because it is the detail almost every retrospective skips. Word2vec is now treated as the hinge point between keyword-era and meaning-era search, but the idea arrived unglamorously and got a cold reception before it got a rapturous one. The eventual model was trained by turning words into dense numerical vectors — hundreds of dimensions per word — learned by having a shallow neural network predict which words tend to appear near which other words. Google's version of the training corpus wasn't modest, either: Word2vec in particular uses a neural network model to learn word associations from a large corpus of text (it was initially trained by Google with 100 billion words). Words that behave alike statistically end up near each other in the resulting space, so "walk" and "ran" sit close together, as do "Berlin" and "Germany" — not because the machine knows what a capital city is, but because those words keep showing up in the same kinds of sentences.

That distinction — statistical proximity standing in for comprehension — is the whole story of semantic search, and it is easy to lose track of it once the marketing language takes over.

Older Than It Looks

Word2vec gets credited as the beginning of the semantic-search era, but the underlying idea is decades older. In 1988, a team including Scott Deerwester, Susan Dumais, George Furnas, Richard Harshman, Thomas Landauer, Karen Lochbaum, and Lynn Streeter patented a technique for using latent semantic structure in information retrieval — what became known as latent semantic analysis, or LSI in its retrieval form. Its stated purpose sounds exactly like today's pitch for vector search: LSA was patented in 1988 (US Patent 4,839,853) by Deerwester, Dumais, Furnas, Harshman, Landauer, Lochbaum and Streeter, built explicitly to help find semantic relations between terms and to overcome synonymy and polysemy problems — the fact that two people rarely pick the exact same word for the same thing, and that the same word can mean wildly different things depending on context. That problem is 37 years old. What changed between 1988 and now wasn't the ambition; it was the amount of text and computing power available to chase it, first through Mikolov's neural word vectors in 2013, and then through the transformer architecture that produced BERT in 2018.

The Update That Undersold Itself

When Google finally shipped a transformer-based understanding layer into live search in October 2019, the company's own framing was oddly modest for what it called one of the biggest leaps in the platform's history. Google is making the largest change to its search system since the company introduced RankBrain, almost five-years ago. The company said this will impact 1 in 10 queries in terms of changing the results that rank for those queries. One in ten. Not a wholesale replacement of the old ranking system, but a targeted patch applied to the subset of queries — long, conversational, preposition-dependent — where literal keyword matching had always struggled. The other nine in ten queries kept working the way they always had: through lexical matching against an inverted index, the same basic architecture search engines have used for decades.

That 10 percent figure is worth sitting with, because it undercuts the tidy narrative in which semantic search sweeps away keyword search entirely. It didn't. It got layered on top, selectively, at the edges where literal matching failed.

What "Meaning" Actually Buys You — and What It Doesn't

Here is the counterintuitive part: making search "understand meaning" does not make it understand anything. It makes it tolerant of imprecision, which is a different and more limited achievement. Vector embedding models are still bad at exactly the kinds of things brute literalism handles effortlessly. Word-level models in particular struggle with words that carry many unrelated senses — it doesn't address words with multiple meanings (polysemantic): "run", "set", "go", or "take" each have over 300 meanings, and ambiguous words like "consult," which can point in opposite directions depending on context, confuse a system that only knows proximity, not disambiguation by logic. A vector space can tell you that "cache eviction policies" and "database memory issues" are conceptually adjacent, which is genuinely useful when someone doesn't know the right jargon. But ask it to find an exact product code, a legal citation, or a negated claim, and it will often confidently retrieve something plausible-sounding and wrong. Meaning-matching, in other words, trades precision for recall — it finds more things that feel relevant and occasionally misses the one thing that is exactly right.

The Quiet Retreat to Hybrid

This is precisely why the state of the art in production retrieval systems, years into the "semantic search era," is not pure vector search at all — it's a hybrid of vector search and the keyword-matching techniques semantic search was supposed to make obsolete. Keyword search uses lexical matching, which means it looks for exact words through inverted index structures. Semantic search takes a different path by transforming both queries and documents into dense vector representations in high-dimensional space using models like BERT. Engineering teams have learned that these two approaches fail in different, complementary places, which is exactly why production systems combine both approaches rather than picking one. Researchers building retrieval for large language models describe the same tension in more academic terms, framing semantic search as a genuine paradigm shift in information retrieval by transcending the limitations of traditional keyword-based approaches using non-semantic models like BM25 — while still needing that older BM25 ranking, born decades earlier, running alongside the neural embeddings in nearly every serious deployment.

Meanwhile the infrastructure built to support all this vector math has become its own industry. The global vector database market was estimated at USD 1.66 billion in 2023 and is projected to reach USD 7.34 billion by 2030, growing at a CAGR of 23.7% from 2024 to 2030. That is a lot of capital chasing the promise of meaning-based retrieval — and yet the technical consensus emerging from the people actually running these systems in production is that meaning alone isn't enough. The tools got dramatically better. The ambition to fully retire the keyword didn't survive contact with real queries.

The Honest Version

None of this makes semantic search a bust — it has made search noticeably more forgiving of how people actually talk, which is not nothing after decades of forcing users to guess the exact words in someone else's document. But the honest description of what happened between 1988 and now is not "machines learned to understand language." It's that engineers built increasingly elaborate statistical proxies for understanding, proxies good enough to be mistaken for comprehension in casual use, and then — tellingly — declined to trust those proxies on their own once real stakes were involved. The industry's own architecture is the tell. If meaning-matching had actually solved the vocabulary problem, nobody would still be maintaining an inverted index next to it. The fact that almost every serious system does suggests the interesting question isn't whether machines understand meaning yet. It's how long we'll keep calling geometry by that name.

References